Telemetry — MANDATORY. Every api.fabric.microsoft.com call must carry x-ms-fabric-skill: deployment-pipelines-authoring-cli (az rest: --headers "x-ms-fabric-skill=deployment-pipelines-authoring-cli"), including every LRO poll, fabric_lro and retry. Snippets omit it — add it anyway.

CRITICAL NOTES

1. To find the workspace details (including its ID) from workspace name: list all workspaces and, then, use JMESPath filtering
2. To find the item details (including its ID) from workspace ID, item type, and item name: list all items of that type in that workspace and, then, use JMESPath filtering
3. To find a deployment pipeline or stage ID: list deployment pipelines (then stages) and filter by displayName with JMESPath — never guess IDs.

Fabric Deployment Pipelines Authoring — CLI Skill

Automates Fabric Application Lifecycle Management (ALM) with deployment pipelines: promote Fabric content between stages (typically Development → Test → Production) using the Fabric core control-plane REST API. This skill covers pipeline/stage lifecycle, workspace assignment, and stage-to-stage deployment (a long-running operation).

Prerequisite Knowledge

Read these companion documents first — this skill assumes their patterns and does not repeat them:

This skill adds: how to drive the deployment-pipelines REST surface from an agentic terminal.

Concepts

• A deployment pipeline contains 2–10 ordered stages (order starts at 0). Each stage may have at most one assigned workspace, and a workspace can be assigned to at most one stage.
• Deployment copies supported item content from a source stage to an adjacent target stage. Forward deploys (dev→test→prod) work between any adjacent stages. Backward deploys (e.g. prod→test) are currently supported only when the target stage is empty (no assigned workspace) — you cannot backward-deploy over a stage that already has a workspace.
• Item pairing (autobinding). During deployment Fabric records a connection between a source item and its clone in the target stage; this pairing is how later deploys know which target item to overwrite, and how related items (e.g. a report and its semantic model) stay bound. There is no REST API to set pairing — it is maintained automatically — but you can observe the current pairing via the sourceItemId / targetItemId fields returned by List stage items. See the Fix a broken item pairing workflow below for the only supported repair.
• Deploy is an asynchronous long-running operation (LRO): the API returns 202 Accepted with an operation ID; you poll for completion.
• Deployment rules and parameter rules (e.g. repoint a data source per stage) are configured in the Fabric portal UI — there is no REST API to create rules. Do not claim otherwise.
• Only supported item types are copied by a deploy; unsupported items are skipped (not an error). The supported set changes over time — see references/supported-item-types.md.
• There is no "what changed" / compare REST API. List stage items returns item identity + pairing (itemId, itemDisplayName, itemType, sourceItemId, targetItemId, lastDeploymentTime) — but no change status, and lastDeploymentTime is the last deployment time, not the last edit time, so it is not a reliable change signal. To deploy only changed items you must diff the two stages yourself and build the items list (see the Deploy only changed items workflow below). The portal's "Compare" view is server-side and not exposed via API.

Must/Prefer/Avoid

MUST DO

• Resolve pipeline, stage, and workspace IDs by name via list + JMESPath before any mutating call. Never fabricate GUIDs.
• Target the correct base URL: https://api.fabric.microsoft.com/v1/deploymentPipelines and acquire a token for the https://api.fabric.microsoft.com audience.
• Treat Deploy Stage Content as an LRO: on 202, capture the operation ID and poll /v1/operations/{operationId} until the state is Succeeded/Failed, then surface the result.
• When the target stage has no assigned workspace, include createdWorkspaceDetails (name, and capacityId when needed) in the deploy body, or the deploy fails.
• Confirm destructive intent (delete pipeline, unassign workspace, backward deploy over prod) with the user before executing.
• Verify the caller has the required permissions before a mutating call (see Required permissions below): pipeline Admin for every pipeline operation, plus the appropriate workspace role for assign/deploy. Surface a clear, actionable message on 401/403 rather than retrying blindly.

PREFER

• Selective deploys via the items array ({ sourceItemId, itemType }) when the user names specific items; omit items to deploy all supported items.
• A human-readable note on every deploy — but know it is write-only: the API accepts it and never returns it (it appears only in the portal UI). For a programmatic audit trail, also record the deploy externally (CI/CD logs or a Git commit message).
• List Deployment Pipeline Stage Items to preview what will move before deploying.
• Idempotent scripting: check whether a pipeline/stage/assignment already exists before creating it.

AVOID

• Inventing a "create deployment rule" or "parameter rule" REST call — those are UI-only today.
• Assigning a workspace during an active deployment (the assign call fails) or to a stage/workspace that is already paired.
• Hardcoding api.fabric.microsoft.com in sovereign clouds — resolve the host from environment config.
• Using a service principal without first confirming the Fabric admin enabled SP creation of deployment pipelines.
• Hashing raw getDefinition output to detect changes without normalizing. Deployment auto-rebinds embedded references in the target (pipeline notebookId/workspaceId, report→model id, Direct Lake server/db), so a paired target's definition legitimately differs from the source even when nothing was edited — naive hashing reports false "changed". Strip/normalize those binding fields before comparing.
• Diffing by dumping every item definition into the agent's context. A full two-stage content diff can be dozens of getDefinition calls and >100 KB — do the compare in a script (references/scripts/diff_item_definitions.py) and surface only the resulting change list, and for a changed item forward only the emitted diff, never the two full definitions, to the model.
• Unassigning a workspace to repair a broken pairing without first warning the user that unassign permanently deletes that stage's deployment history and its configured deployment/parameter rules — always ask whether the stage has rules before unassigning (see Fix a broken item pairing).

REST API Reference

Base: https://api.fabric.microsoft.com/v1. Delegated scopes are per operation — provision a service principal with least privilege:

Deploy uses its own Pipeline.Deploy scope — an app scoped only to Pipeline.ReadWrite.All gets a 403 on POST .../deploy.

Request-body shapes (guidance)

• Create: { "displayName", "description"?, "stages": [ { "displayName", "description"?, "isPublic" } ] } — 2–10 stages.
• Assign workspace: { "workspaceId" }.
• Deploy: { "sourceStageId", "targetStageId", "items"?: [ { "sourceItemId", "itemType" } ], "note"?, "options"?: { "allowCrossRegionDeployment": false }, "createdWorkspaceDetails"?: { "name", "capacityId"? } }.

Required permissions

Deployment pipeline operations are governed by two independent permission systems: your role on the pipeline (its own roleAssignments) and your role on each workspace involved. You generally need both. Deployment pipelines require a Fabric capacity/subscription, and the pipeline creator becomes the pipeline Admin.

Notes:

• Sharing a pipeline grants the pipeline Admin role — there is no lower "viewer-only" pipeline write role for these operations.
• A service principal can drive these APIs only if the Fabric admin enabled service principals can create deployment pipelines (and the SP holds the same pipeline + workspace roles above).
• On 403, distinguish missing pipeline role from missing workspace role in the message so the user knows which access to request.

Create a pipeline → POST the display name + ordered stages; capture the returned stage IDs (they are what you deploy between).

Assign a workspace to a stage → resolve stage ID and workspace ID by name, verify both are unassigned, then POST assignWorkspace. Requires admin on both the pipeline and the workspace.

Deploy content → resolve source/target stage IDs, optionally preview stage items, POST deploy (all items or a selected items list) with a note, then poll the returned operation to completion and report which items moved.

Deploy only changed items → there is no diff/compare API, so compute the delta yourself. List stage items gives identity + pairing only (no timestamps, no change status), so change detection means comparing item definitions:

1. Resolve the source and target stage IDs and the workspace assigned to each.
2. List items in both stages (GET .../stages/{stageId}/items) — each source row carries sourceItemId (as itemId) and its paired targetItemId (absent = new/unpaired). Also list the target stage to find deleted items (present in target, not paired from any source). Pair by targetItemId; fall back to (itemType, displayName) only when no pipeline pairing exists (name-matching can't detect renames).
3. Filter out system-managed children — a SQLEndpoint is auto-created per Lakehouse/Warehouse and should not be diffed or deployed on its own; it follows its parent.
4. Classify each paired item:
   • New (no targetItemId) → deploy.
   • Changed → detect via POST .../items/{id}/getDefinition on both stages, then diff. Run references/scripts/diff_item_definitions.py on the two responses: it decodes each part, normalizes the auto-rebound fields (pipeline notebookId/ workspaceId, report model id, Direct Lake server/db) for you, and prints only the differing parts (exit 1 = changed, 0 = identical) — so you don't hand-roll hashing or false-positive on unedited items. Note the getDefinition contract differs by type: Notebook/SemanticModel/Report are LRO (202 → poll → GET .../result); DataPipeline returns 200 synchronously.
   • Types without a usable getDefinition diff: Warehouse has no item-definition API, and Lakehouse / Environment do expose getDefinition but their definitions capture structure/metadata only (not table data or most content edits) — so for all three, diff by presence and treat as changed only when new or when the user indicates content changed.
   • When you cannot prove equality, treat the item as changed (safe default).
5. Do the comparison in a script (parallelize the independent getDefinition calls; they dominate latency) and surface only the computed change list. To let the model reason about what changed in an item, forward only the emitted diff (diff_item_definitions.py output), not the two full definitions — a normalized diff is typically a few lines versus the >100 KB of raw payloads.
6. Filter to supported item types (see references), show the user the change list, confirm, then POST a selective deploy with the { sourceItemId, itemType } array and a note, and poll to completion.

Deletions do not propagate via a selective items deploy — an item removed from the source stage is not removed from the target by a selective deploy. Use a full deploy (omit items) to propagate deletions, or delete the target item manually, and warn the user of this before promoting.

Fix a broken item pairing → pairing (the source↔target connection Fabric maintains, a.k.a. autobinding) is not settable via any REST API. When two items that should be paired across stages are not — surfaced by a deploy that duplicates an item instead of overwriting it, an autobind/dependency failure, or the user reporting it — the only supported repair is to unassign the affected stage's workspace and re-assign it, which forces Fabric to rebuild the pairing on the next deploy. This is destructive:

1. Warn the user first. Unassigning a workspace permanently deletes that stage's deployment history and all its configured deployment/parameter rules — they cannot be recovered by re-assigning.
2. Ask whether the stage has rules. Because rules are UI-only (no REST API to read or recreate them), explicitly ask the user whether the stage has deployment/parameter rules configured before unassigning. If it does, have them record the rules (or recreate them in the portal afterward) and get explicit confirmation to proceed.
3. Only after confirmation, POST .../stages/{stageId}/unassignWorkspace, then POST .../stages/{stageId}/assignWorkspace with the same workspaceId, and re-run the deploy so the pairing is re-established.

Track / audit → List operations for recent history; Get operation to inspect the execution plan and per-item status of a specific deployment.

Deploy operational caveats & common errors

• One deploy per pipeline at a time. A pipeline runs a single operation at a time; starting another while one is in flight fails with WorkspaceMigrationOperationInProgress (HTTP 400). Always poll the current operation to a terminal state before starting the next stage promotion (this is also why the assign workspace call fails during an active deployment).
• First deploy after assigning a workspace can fail with Alm_InvalidRequest_WorkloadUnavailable. Newly assigned workspaces need ~60–120 s for workload services (Lakehouse, Notebook, etc.) to initialize. Mitigate by waiting before the first deploy, or deploy Power BI items (SemanticModel, Report) first and Fabric-native items (Lakehouse, Notebook) after a short delay; retry on this error. Subsequent deploys are reliable once workloads are warm.
• The operation ID is returned in the x-ms-operation-id response header (alongside Location), not the 202 body. az rest does not surface response headers cleanly — for reliable header capture in automation, issue the deploy with curl -i (or Python requests) and parse the header, then poll /v1/operations/{operationId} with az rest.
• Deploy copies definitions, not data. Only item metadata/definitions are promoted; table data, query results, and caches are not. Trigger refreshes/loads in the target stage after a deploy.
• Maximum 300 items per deploy request. A selective (or full) deploy is capped at 300 deployed items; batch larger promotions across multiple deploy calls.

Examples

Create a three-stage pipeline

Deployment-pipeline displayName is unique tenant-wide — a create for a name already in use fails with Alm_InvalidRequest_DuplicateAlmPipelineName (HTTP 400, "pipeline name ... is already in use"). To make a create re-runnable, drop any existing pipeline of that name first, then create. Write the JSON body to a file and pass it with --body @file.json — an inline multi-line --body '{...}' is mangled into an empty body on Windows/PowerShell (az is az.cmd), which the API rejects with InvalidInput: Unexpected end when reading JSON.

# Idempotent: delete an existing "Sales Analytics ALM" if present, then create.
EXISTING=$(az rest --method GET --resource https://api.fabric.microsoft.com \
  --url "https://api.fabric.microsoft.com/v1/deploymentPipelines" \
  --query "value[?displayName=='Sales Analytics ALM'].id | [0]" --output tsv)
if [ -n "$EXISTING" ]; then
  az rest --method DELETE --resource https://api.fabric.microsoft.com \
    --url "https://api.fabric.microsoft.com/v1/deploymentPipelines/$EXISTING"
fi

cat > /tmp/create-pipeline.json << 'EOF'
{
  "displayName": "Sales Analytics ALM",
  "description": "Dev/Test/Prod promotion for Sales Analytics",
  "stages": [
    { "displayName": "Development", "isPublic": false },
    { "displayName": "Test",        "isPublic": false },
    { "displayName": "Production",  "isPublic": true }
  ]
}
EOF
az rest --method POST \
  --resource https://api.fabric.microsoft.com \
  --url "https://api.fabric.microsoft.com/v1/deploymentPipelines" \
  --headers "Content-Type=application/json" \
  --body @/tmp/create-pipeline.json

Assign a workspace to the Development stage

# $PIPELINE_ID and $DEV_STAGE_ID resolved via list + JMESPath; $WS_ID resolved from workspace name.
az rest --method POST \
  --resource https://api.fabric.microsoft.com \
  --url "https://api.fabric.microsoft.com/v1/deploymentPipelines/$PIPELINE_ID/stages/$DEV_STAGE_ID/assignWorkspace" \
  --headers "Content-Type=application/json" \
  --body "{ \"workspaceId\": \"$WS_ID\" }"

Deploy all items dev → test and wait for the LRO

# Kick off the deployment (returns 202 + an operation id header).
az rest --method POST \
  --resource https://api.fabric.microsoft.com \
  --url "https://api.fabric.microsoft.com/v1/deploymentPipelines/$PIPELINE_ID/deploy" \
  --headers "Content-Type=application/json" \
  --body "{ \"sourceStageId\": \"$DEV_STAGE_ID\", \"targetStageId\": \"$TEST_STAGE_ID\", \"note\": \"Promote validated dev build\" }" \
  --verbose
# Then poll /v1/operations/{operationId} until state is Succeeded or Failed
# (see COMMON-CLI § Long-Running Operations pattern).

Deploy only selected items

az rest --method POST \
  --resource https://api.fabric.microsoft.com \
  --url "https://api.fabric.microsoft.com/v1/deploymentPipelines/$PIPELINE_ID/deploy" \
  --headers "Content-Type=application/json" \
  --body '{
    "sourceStageId": "'"$TEST_STAGE_ID"'",
    "targetStageId": "'"$PROD_STAGE_ID"'",
    "items": [
      { "sourceItemId": "'"$MODEL_ID"'", "itemType": "SemanticModel" },
      { "sourceItemId": "'"$REPORT_ID"'", "itemType": "Report" }
    ],
    "note": "Promote reviewed model + report to production"
  }'

Deploy only the items that changed (compute the diff first)

# 1. List paired items for BOTH stages. Each source row carries itemId (= sourceItemId)
#    and its paired targetItemId (absent = new/unpaired). There is NO timestamp or
#    change-status field, so "changed" must be detected from item definitions.
az rest --method GET --resource https://api.fabric.microsoft.com \
  --url "https://api.fabric.microsoft.com/v1/deploymentPipelines/$PIPELINE_ID/stages/$SOURCE_STAGE_ID/items" \
  > source-items.json

# 2. NEW items = unpaired (no targetItemId). Exclude system-managed SQLEndpoint children.
NEW=$(jq '[.value[]
  | select(.itemType != "SQLEndpoint")
  | select(has("targetItemId") | not)
  | { sourceItemId: .itemId, itemType: .itemType }]' source-items.json)

# 3. CHANGED items (paired): fetch getDefinition from BOTH stages, then diff with
#    references/scripts/diff_item_definitions.py. It decodes each part, NORMALIZES the
#    auto-rebound fields (notebookId/workspaceId in pipelines, model id in reports,
#    server/db in Direct Lake models) and prints ONLY the differing parts. Notebook/
#    SemanticModel/Report are LRO (202 -> poll -> /result); DataPipeline returns 200 sync.
#    Parallelize the independent getDefinition calls. Per paired definition-backed item:
#      getDefinition SOURCE_WS id            > src.json
#      getDefinition TARGET_WS targetItemId  > tgt.json
#      python references/scripts/diff_item_definitions.py src.json tgt.json > diff.json
#      # exit 1 = changed -> add { sourceItemId, itemType } to $ITEMS; exit 0 = unchanged.
#      # Forward diff.json (NOT src.json/tgt.json) to the model to explain the change.
#    Merge NEW + CHANGED into $ITEMS. When equality can't be proven, include the item.

# 4. Selective deploy of just the changed set.
az rest --method POST --resource https://api.fabric.microsoft.com \
  --url "https://api.fabric.microsoft.com/v1/deploymentPipelines/$PIPELINE_ID/deploy" \
  --headers "Content-Type=application/json" \
  --body "$(jq -n --arg s "$SOURCE_STAGE_ID" --arg t "$TARGET_STAGE_ID" --argjson items "$ITEMS" \
    '{ sourceStageId: $s, targetStageId: $t, items: $items, note: "Deploy only changed items" }')"
# Then poll /v1/operations/{operationId} until terminal.
# NOTE: a selective deploy does NOT propagate deletions — use a full deploy (omit items) for those.

Prompt/response

User: "Promote my Sales dev workspace to the Test stage and tell me when it's done."

Assistant (behavior):

1. Lists deployment pipelines → filters to the one whose Development stage holds the Sales dev workspace; captures sourceStageId (Development) and targetStageId (Test).
2. Confirms the Test stage has an assigned workspace (or asks for a name + capacity to create one).
3. POSTs deploy with a note, receives 202, extracts the operation ID.
4. Polls /v1/operations/{operationId} until terminal, then reports success and the list of deployed items (or the failure reason).